Designer Perspective

From a designer's perspective, the features that the UIA requires to exhibit the behaviors discussed above are listed below:

• Knowing The UIA has to know how to interpret user instructions and translate them into actions that it can perform. For this, it needs knowledge about the entities and their relationships for the given TOS. That is, it needs to know about the valid objects, properties, values, ranges, types, tasks, commands, readings, names, relationships between objects (e.g., an object may form a category of certain other objects), relationships between commands (e.g., one command may start a task or a reading and another command may stop a task or a reading), relationships between names (e.g., two names may be equivalent) and relationships between tasks and readings (e.g., certain readings may be the effects of certain tasks).

  In addition, the UIA needs to have knowledge about when different actions are to be performed. This requires knowledge about time and its passage.

  Since an agent's knowledge can change (either because of connecting to a new TOS or because of the dynamic nature of the environment), the knowledge base of the agent should be easily modifiable, so that new information can be added and existing information can be updated or deleted relatively quickly.

• Reasoning The agent has to not only know things, but through reasoning, it has to take explicit knowledge and go beyond it to implicit knowledge that can be derived from what is known as well as make decisions or perform actions based on the knowledge. Not only that, since the knowledge that an agent has may be false, imperfect or even contradictory, the reasoning has to be robust enough to deal with contradictory information and changing beliefs.

  The agent has to perform theoretical reasoning directed towards determining the beliefs of the agent as well as practical reasoning [Dav01,Woo00] that is directed towards determining the actions that the agent can perform. Theoretical reasoning involves creating new beliefs based on the current beliefs, as in deriving the new belief ``Water is warm'', from the beliefs ``The temperature of water is $100^oF$'' and ``Any water with temperature between $95^oF$ and $110^oF$ is warm''. On the other hand, practical reasoning involves deciding what desires are to be pursued and how to pursue them.

  By reasoning with its knowledge, the agent has to interpret each user instruction and then translate it into appropriate UIA actions. For instance, if a user instruction is ``Switch on the heater whenever the pool temperature drops below $75^oF$'', then using its reasoning capability, an agent should be able to send ``READ TEMPERATURE'' periodically and issue ``HEAT POOL'' command at appropriate times without any further intervention from the user.

• Meta-reasoning Meta-reasoning can help identify any anomalies or inconsistencies in the regular reasoning process, thus providing some level of perturbation tolerance for the resulting agent. This can be illustrated with an example. Suppose, the user initially gives the instruction ``Switch on the heater at $11:00 am$ on non-working days if the temperature is below $75^oF$''. The reasoning process could decide when the ``HEAT POOL'' commands need to be issued based on the result of ``READ TEMPERATURE'' command and the knowledge about the user's work schedule. Now, suppose the user gives a new instruction ``Do not heat the pool during my vacation days'' before going on a vacation. Assuming that the agent has the background knowledge that the user does not work during vacation days, these two instructions can cause an inconsistency in the agent's belief set since the first instruction causes it to believe that ``HEAT POOL'' has to be issued, while the latter instruction causes it to believe that ``HEAT POOL'' should not be issued during the vacation days. In such scenarios, meta-reasoning can help the reasoning process make a decision regarding how to proceed. In the current example, one possibility is to note that ``vacation days'' is a subset of non-working days and hence give more importance to the belief associated with the more specific subset (vacation days) than the more general class (non-working days), thus deciding not to sent the ``HEAT POOL'' command after all. Another heuristic that can be applied to deal with inconsistency is to give more importance to a user's latter instruction. Thus, meta-reasoning can provide the agent the ability to go beyond reacting to changes in the environment to actually taking the initiative under specific circumstances, thereby exhibiting some level of proactivity.